From 0ae4740bd111ce3197baf7aa836b4224a0b4c4d0 Mon Sep 17 00:00:00 2001
From: Qeeweew <68716978+Qeeweew@users.noreply.github.com>
Date: Wed, 17 Jun 2026 10:29:57 +0800
Subject: [PATCH] fix: add missing clamp_limit for CompressedTensorsWNA16MoE
 (#27328)

---
 .../compressed_tensors/schemes/compressed_tensors_wNa16_moe.py   | 1 +
 1 file changed, 1 insertion(+)

diff --git a/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16_moe.py b/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16_moe.py
index 581972032..0d6a9b896 100644
--- a/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16_moe.py
+++ b/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16_moe.py
@@ -425,6 +425,7 @@ class CompressedTensorsWNA16MoE(CompressedTensorsMoEScheme):
             num_bits=self.num_bits,
             is_k_full=self.is_k_full,
             routed_scaling_factor=self.moe_runner_config.routed_scaling_factor,
+            clamp_limit=self.moe_runner_config.swiglu_limit,
             workspace=layer.workspace,
         )
         return StandardCombineInput(hidden_states=output)